【摘要】 以传统的U型网络为架构，提出一种基于多尺度特征提取与多特征融合的甲状腺结节超声影像分割方法。首先，设计一种基于多个小卷积核叠加的特征提取策略。通过堆叠多个小尺寸的卷积核，模型能够在不同的感受野下捕捉图像中的细节特征和全局特征，从而实现多尺度特征的高效提取。其次，通过混合注意力机制，包括通道注意力与空间注意力，将不同阶段的特征图进行融合，以强化原有的跳跃连接。本文算法在甲状腺结节分割数据集TN3K和DDTI的95%豪斯多夫距离（HD95）分别为16.02和17.86 mm,F1分数分别为82.21%和75.74%，在所有对比算法中表现最佳。实验结果表明，该方法可以为临床医生提供辅助诊断。更多还原

【Abstract】 Based on the classical U-Net, a novel method incorporating multi-scale feature extraction and multi-feature fusion for thyroid nodule segmentation in ultrasound image is proposed. Specifically, a feature extraction strategy based on stacked small-sized convolutional kernels is designed. By stacking multiple small-sized convolutional kernels, the model can capture both detailed and global features of images under different receptive fields, thereby achieving efficient multi-scale feature extraction. Through a hybrid attention mechanism which includes both channel and spatial attention, feature maps from different stages are effectively fused, thereby enhancing the original skip connections. The proposed algorithm achieves 95%Hausdorff distance(HD95) of 16.02 and 17.86 mm on the TN3K and DDTI thyroid nodule segmentation datasets,respectively, along with F1-scores of 82.21% and 75.74%, outperforming all other compared methods. Experimental results demonstrate that this approach can provide valuable assistance to clinicians in diagnostic practice.更多还原